Update README.md
Browse files
README.md
CHANGED
|
@@ -37,8 +37,14 @@ model = ClassificationModel("deberta", "microsoft/deberta-large", args=model_arg
|
|
| 37 |
model.train_model(train_df, eval_df=eval_df)
|
| 38 |
```
|
| 39 |
|
|
|
|
|
|
|
| 40 |
**Note**: This dataset is highly imbalanced and it is recommended to use a library like [imbalanced-learn](https://imbalanced-learn.org/stable/) before proceeding with training.
|
| 41 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 42 |
## Feature description
|
| 43 |
|
| 44 |
- `text_a`: The utterance prior to the utterance being classified. (Say for dialog with turns 1-2-3, if we are trying to find the dialog act for 2, text_a is 1)
|
|
|
|
| 37 |
model.train_model(train_df, eval_df=eval_df)
|
| 38 |
```
|
| 39 |
|
| 40 |
+
## Balanced variant of the training set
|
| 41 |
+
|
| 42 |
**Note**: This dataset is highly imbalanced and it is recommended to use a library like [imbalanced-learn](https://imbalanced-learn.org/stable/) before proceeding with training.
|
| 43 |
|
| 44 |
+
Since, balancing can be complicated and resource-intensive, we have shared a balanced variant of the train set that was created via oversampling using the _imbalanced-learn_ library. The balancing used the `SMOTEN` algorithm to deal with categorical data clustering and was resampled on a 16-core, 60GB RAM machine. You can access it using:
|
| 45 |
+
|
| 46 |
+
```load_dataset("diwank/silicone-merged", "balanced")```
|
| 47 |
+
|
| 48 |
## Feature description
|
| 49 |
|
| 50 |
- `text_a`: The utterance prior to the utterance being classified. (Say for dialog with turns 1-2-3, if we are trying to find the dialog act for 2, text_a is 1)
|